Llama 3.1 Megathread - eviltoast

Meta has released and open-sourced Llama 3.1 in three different sizes: 8B, 70B, and 405B

This new Llama iteration and update brings state-of-the-art performance to open-source ecosystems.

If you’ve had a chance to use Llama 3.1 in any of its variants - let us know how you like it and what you’re using it for in the comments below!

Llama 3.1 Megathread

For this release, we evaluated performance on over 150 benchmark datasets that span a wide range of languages. In addition, we performed extensive human evaluations that compare Llama 3.1 with competing models in real-world scenarios. Our experimental evaluation suggests that our flagship model is competitive with leading foundation models across a range of tasks, including GPT-4, GPT-4o, and Claude 3.5 Sonnet. Additionally, our smaller models are competitive with closed and open models that have a similar number of parameters.

As our largest model yet, training Llama 3.1 405B on over 15 trillion tokens was a major challenge. To enable training runs at this scale and achieve the results we have in a reasonable amount of time, we significantly optimized our full training stack and pushed our model training to over 16 thousand H100 GPUs, making the 405B the first Llama model trained at this scale.


Official Meta News & Documentation

See also: The Llama 3 Herd of Models paper here:


HuggingFace Download Links

8B

Meta-Llama-3.1-8B

Meta-Llama-3.1-8B-Instruct

Llama-Guard-3-8B

Llama-Guard-3-8B-INT8


70B

Meta-Llama-3.1-70B

Meta-Llama-3.1-70B-Instruct


405B

Meta-Llama-3.1-405B-FP8

Meta-Llama-3.1-405B-Instruct-FP8

Meta-Llama-3.1-405B

Meta-Llama-3.1-405B-Instruct


Getting the models

You can download the models directly from Meta or one of our download partners: Hugging Face or Kaggle.

Alternatively, you can work with ecosystem partners to access the models through the services they provide. This approach can be especially useful if you want to work with the Llama 3.1 405B model.

Note: Llama 3.1 405B requires significant storage and computational resources, occupying approximately 750GB of disk storage space and necessitating two nodes on MP16 for inferencing.

Learn more at:


Running the models

Linux

Windows

Mac

Cloud


More guides and resources

How-to Fine-tune Llama 3.1 models

Quantizing Llama 3.1 models

Prompting Llama 3.1 models

Llama 3.1 recipes


YouTube media

Rowan Cheung - Mark Zuckerberg on Llama 3.1, Open Source, AI Agents, Safety, and more

Matthew Berman - BREAKING: LLaMA 405b is here! Open-source is now FRONTIER!

Wes Roth - Zuckerberg goes SCORCHED EARTH… Llama 3.1 BREAKS the “AGI Industry”*

1littlecoder - How to DOWNLOAD Llama 3.1 LLMs

Bloomberg - Inside Mark Zuckerberg’s AI Era | The Circuit

  • brucethemoose@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    ·
    3 months ago

    Yeah, well, I have been using base models and a few instruct tunes for a bit and haven’t even gotten refusals, as long as there as enough existing context.

    • hendrik@palaver.p3x.de
      link
      fedilink
      English
      arrow-up
      1
      ·
      edit-2
      3 months ago

      Sure. Depends entirely on which instruct tuned model you choose. The official one from Meta has guardrails in place. A community tuned one (like Fimbulvetr, Stheno, …) generally build on the base variant and comes without any guardrails. There are exceptions to the rule, some extend the other variant or some other model. I think Mistral also has censorship. And the big ones like ChatGPT are heavily censored. This also makes it creep into other models if they heavily rely on synthetic data from ChatGPT without mitigating for that. I’m not sure which model you took.

      I really dislike being lectured by ChatGPT, being forced to have an argument with my computer before it does the task I hand out. It’s not a big deal for some tasks like enhancing my mails, but I’m trying to get a bit more creative with AI. And things I really had difficulties with were writing a detective / murder mystery story, writing (lewd) fan fiction and songtexts which I want to generate for Suno. I just avoid all the “censored” models and ChatGPT, because it usually goes on a long tangent and lectures me about murder and why it’s wrong and how you have to consider both sides of a story. Plain refuses to talk about some stuff and won’t touch adult topics. Even if I tell it 10 times that it’s pretty common to write songs or stories about such things.
      I’ve tried several “jailbreaks” and maybe I’m too late to the game, because I found dozens of prompts for ChatGPT and not even one worked for me. (And these are only some examples. I think these tools are also not supposed to give you medical advice, be a lawyer … and I regularly get something trigger the safety mechanisms. But I’m a responsible adult and want to be treated like one.)

      So I just use the Llama models. And I like their tone and way of speaking to me way better. Hence my question if there’s a base model available and I can expect some good community fine-tunes in the future, entirely without refusals and long lectures about ethics. And ideally I’d like to not have it censored and the un-censired and use a degraded version, but an unrestricted one from the community like I used before. (Btw also Meta seems to have improved their approach. Also the newer “censored” versions don’t outright refuse. They’re very polite and try to offer you an alternative. But sadly I want what I asked for and not something else.)

      I think I can answer my question now. There wasn’t much first-hand information available when I first asked. But it seems they have a base variant and an instruct tuned variant available. I’m still not sure if they did something to it, but if they stick to what the term “base model” meant, it should be what I’m looking for.